I build what used to take a team. You get one person who can be reached.

I'm Stéphane Lepain, founder of CPLT. I document the process, choose the workspace and models with you, build it in an environment you own and hand it over. I agree the permitted information, human checks, scope and price before paid work starts. Existing tools are welcome; private hosting is an option. Everything here is built around your workflow. The one exception is a plain chat assistant with no automation, which is as close to off-the-shelf as this gets.

Describe the task, how often it happens and who is responsible. I look at the current process before choosing a tool. You do not need infrastructure knowledge or confidential files for the first conversation.

I can be the first automation engineer you bring in, without creating a permanent internal role. I start by understanding one process and the decisions behind it. I agree with you on where AI helps, which model fits and where a person checks the result. I build around those choices, document how it runs and walk your team through the handover so your people can operate it.

Read what an automation engineer does, and when a scoped engagement makes sense before deciding whether to create a permanent role.

I can build an assistant that answers questions from your own documents and shows the source page for each answer. I agree the information it can access and test representative questions with you. Your team checks the answers before relying on them.

An answer needs checking against its source. Missing information, wrong citations and misunderstood documents are cases to test, not reasons to trust a fluent response.

We can start where the work repeats most. Read how to write acceptance checks for an internal knowledge assistant before choosing which questions it should handle.

I scope and build workflows that move information through agreed steps, with checks and human approval where needed. For example, a fictional workflow might receive a request, extract details, prepare a draft and send missing information to a reviewer.

This is an illustration, not a ready-made integration or customer result. I verify each proposed connection, permission and action before including it in the written scope. Important decisions remain with the named person responsible.

See a fictional source check and what it does not prove.

I start with a free 45-minute remote scoping call and a one-page written note on fit and next steps. If a trial is needed, it is a separate paid engagement with a written scope and a price agreed first.

• One task, representative inputs, formats, volumes and any connections.
• Expected results, review responsibility and cases to reject or leave unresolved.
• Limits on delivery effort and revisions, with acceptance checks.
• A comparison of current work and trial results at the same quality standard.
• A decision to continue, revise or stop.

I count implementation, model/API usage, hosting, human review, correction, maintenance and support. That includes unsuccessful attempts and work that remains manual.

Released time is capacity. Cash is saved only when spending actually falls. I do not promise a saving, payback period or reliability rate.

Read how to measure an AI pilot's cost per accepted result to define the comparison before committing to a trial.

Before choosing tools, use the guide to scoping one recurring task to name the inputs, reviewer, permitted actions and stop conditions.

If your requirements justify it, I build a workspace in your environment. I assess model size, document handling, simultaneous use, operating responsibility and cost before sizing hardware.

Approved external APIs are another option. I document which information may reach each provider and check the configured routes and applicable terms. External APIs use your provider accounts and API keys; consumer chat subscriptions do not cover API usage.

Hosting locally does not, by itself, prove that no data leaves the environment. Applications, models, backups and connected services all need to fit the agreed boundary.

For the cost decision, read my server-versus-API cost comparison. For an external provider, check the data-processing agreement and responsibilities as well as the technical route.

I deliver paid projects with a written scope and a price agreed before work starts. You work directly with me, from scoping to handover.

I do not sell model subscriptions or seats. Provider charges and licences still apply. There is no retainer requirement; ongoing operations and support can be scoped separately.

When is an existing tool enough?

If an existing tool meets your needs, I will say so. A custom build is not a prerequisite for useful assistance. I would check the existing option against four questions first:

• May it use the information needed for the task?
• Can the person responsible check the result?
• Can the team handle the exceptions and any manual steps?
• Can the team operate it at an acceptable full cost?

I would consider a custom workflow only for a gap that matters to the task, then test whether the build actually closes it. Changing the process or keeping a step manual can also be the right decision.

I document configuration, acceptance results and operating procedures, then walk your team through the runbook. For a hosted system, I agree monitoring, backup and recovery requirements and test the relevant procedures in your environment.

The handover needs to identify what to do when I am unavailable. Support scope, hours and price are agreed separately; there is no promise of unlimited support or automatic maintenance.

Before accepting a handover, I would ask the person taking over to use the runbook to complete a few agreed checks:

• Update a permitted source and check the resulting answer.
• Recognise a failed result and send it to the named reviewer.
• Pause the workflow and use the agreed manual alternative.
• Find who owns an expired account or missing permission.

I would agree which checks apply before the build. These are examples, not a claim that every system supports the same actions or that a customer has already passed them.

What happens, how often, who does it and what would improve it? I reply personally within one business day.